PAFFI: Performance Analysis Framework for Fog Infrastructures in - - PowerPoint PPT Presentation

ICCCS - Oct. 10-12, 2019, Rome 1

PAFFI: Performance Analysis Framework for Fog Infrastructures in realistic scenarios

Claudia Canali, Riccardo Lancellotti

Department of Engineering “Enzo Ferrari” University of Modena and Reggio Emilia

ICCCS - Oct. 10-12, 2019, Rome 2

Fog computing

• Cyber-physical systems

– Distributed sensors – → Huge amount of

information to handle

• Cloud approach:

– High latency – Risk of network

congestion

• Some critical applications:

– Autonomous driving – Support for smart cities

• Alternative paradigm

→ Fog computing

– Presence of Fog nodes – Data aggregation and

filtering

– Latency-bound tasks

ICCCS - Oct. 10-12, 2019, Rome 3

Challenges of Fog computing

• Service placement

– Which services on fog / cloud

• Data flows mapping

– Sensor nodes to fog nodes connection

• Adaptive load balancing

– Cooperation strategies

• → Need for realistic scenarios

– Use of geo-referenced data – Flexible generation of experimental setups – Help for performance evaluation

ICCCS - Oct. 10-12, 2019, Rome 4

Introducing PAFFI

• Performance Analysis Framework

for Fog Infrastructures

• Realistic scenarios based on geographic data
• Support for performance analysis

→ OMNeT++ simulation framework

• Plugin-based approach for topology mapping

→ arbitrary connections among nodes

• Highly flexible and configurable

→ Python as main development tool

ICCCS - Oct. 10-12, 2019, Rome 5

Performance model

• Performance based on queuing theory
• 3 types of nodes: sensor, fog, cloud
• Description of node behavior:

– Outgoing data rate from sensor i:λi – Processing rate at fog node j: μj

• Topology connections:

– Sensor → Fog connections: xi,j – Fog → Cloud connections: yj,k – Network delay: δi,j δj,k

ICCCS - Oct. 10-12, 2019, Rome 6

Introducing PAFFI

• Contributions to response time:

– Sensor → Fog avg net delay – Fog → Cloud avg net delay – Fog processing time

• Parameters to describe scenarios

– Avg net delay – Net / Proc balance – System load

ICCCS - Oct. 10-12, 2019, Rome 7

Framework architecture

• 3 main components
• Use of external services

– Nominatim API (Open Street Map) – AMPL optimization language – OMNeT++ simulation framework

ICCCS - Oct. 10-12, 2019, Rome 8

Geo-referencing

• Input: list of POIs
• Output: geo-referenced and validated data

ICCCS - Oct. 10-12, 2019, Rome 9

Scenario generation

• Connect topology:

– Naive connection

(nearest node)

– Optimized connection (AMPL)

• Sub-sample data
• Create scenario

(δi,j δj,k λi μj)

New connection policies can be easily added

ICCCS - Oct. 10-12, 2019, Rome 10

Performance evaluation

• Create OMNeT++ files:

– Simulated network description (.ned) – Simulation parameters (.ini)

• Use of template files

ICCCS - Oct. 10-12, 2019, Rome 11

PAFFI in action

• Use of geo-referenced data:

– Traffic/Air pollution

monitoring in Modena

• Scenario Comparison

– Naive model – Optimized model

• Representation of

sensor → fog mapping

• Uneven distribution of

sensors over fog nodes

Naive Optimized

ICCCS - Oct. 10-12, 2019, Rome 12

Simulations

• Scenario:

– Delay:δμ = 10ms – Net/Proc: δμ = 1 – Load: ρ = 0.5

• Creation of simulation

– Leverage OMNeT++ GUI – Built-in analysis tools

Naive Optimized

ICCCS - Oct. 10-12, 2019, Rome 13

Performance analysis

Parameter Naive Mappping Optimized mapping Utilization 0.30, ~1, 0.075 0.54, 0.53, 0.45 Queue length 0.07, >1987, 0.0031 0.33, 0.32, 0.19 Queuing time [ms] 2.2, >17650, 0.41 6.0, 5.9, 4.1 Response time [ms] >12807 30.8 Queuing time [ms] >12786 5.4 Processing time [ms] 10 10

• Preliminary performance comparison

– Evidence of overload in a fog node for naive

mapping

– No performance degradation when

connections are optimized

Cloud Fog

ICCCS - Oct. 10-12, 2019, Rome 14

Performance analysis

Naive Optimized

ICCCS - Oct. 10-12, 2019, Rome 15

• Focus on naive mapping
• Given #sensors, #fog

– Sensors for each fog node? – Probability distribution

• Analysis:

– Estimate risk of congestion – Create realistic

heterogeneous scenarios → for load sharing

• Just another script

– Create topology – Collect data

Another example

• Observation:

– Distribution: truncated

Gaussian

– Mean and variance can be

quantified

ICCCS - Oct. 10-12, 2019, Rome 16

PAFFI is available now! Find it in my homepage:

http://web.ing.unimo.it/rlancellotti/

Or send me an email: riccardo.lancellotti@unimore.it